What is Excessive Agency?
Excessive agency, told as a story about handing the apprentice the whole knife drawer and letting it take out anything with no approval at all.
The apprentice got the whole drawer opened for it.What used to be approved errand by errand got handed over all at once.
The broader the access, the more a trick can do.The apprentice cannot tell a smuggled note from a real request either.
Excessive agency means giving far more access than the job needs.Almost nothing is locked out of reach.
This is the difference between proper access and excessive agency.The longer the tool list, the more there is to watch.
Permission design is one thing; excessive agency is going past it.Broad access hands broad power to a smuggled note too.
Excessive Agency = giving the apprentice far more access and automatic approval than the job needs.
Excessive agency describes an AI agent holding more permissions, tool access, or approval authority than its task actually requires. Because an agent cannot tell a smuggled instruction from a genuine one just by reading it, the wider its access, the greater the damage when it is fooled.
When grown-ups say it
- Excessive Agency
- Giving the apprentice far more access than the job needs. There is also an intro to this whole world. → The Kitchen That Takes Any Order
- Agent
- The apprentice that plans its own errands once given a goal. This is exactly who ends up with too much access. → The Parrot That Plans Its Own Errands
- Tool Calling
- Taking a tool out of the knife drawer to use it. The underlying AI concept. → The Toolbox
- Per-tool Permission
- Designing a different badge for each drawer slot. Excessive agency is going past this design. → A Permission Badge Per Tool
- Least Privilege
- Granting only the access that is actually needed. Excessive agency is exactly this principle left unobserved. → Only the Keys You Need
- Sandbox
- An empty room where a suspicious box gets opened first. A way to contain the space a tool runs in. → The Empty Room with No Windows
- Human-in-the-Loop
- A person stamping approval at the very end. An important errand needs this check. → A Person at the End
- Observability
- Looking in three ways — dashboards, logs, and the path a request took. Logging every errand belongs here. → Looking In Three Different Ways